NuML_Studio

This is a UI tool for generating TFLM ML model projects and collecting real-time data for Nuvoton ML MCUs. It can be executed on Windows without the need to install Python or any related libraries.

How this works

• The GitHub repository mainly provides source code. For a directly executable version, please download the .7z file from release and extract it to run.

  • If you want to install from source, refer to the "Build yourself" steps below

• Double-click NuML_Studio.exe to launch the tool.

  • If you want to upload to Edge Impulse or download single test data to deployment, please update your project API key in API_key.txt.

• Quick start guide for using NuML_Studio and Edge Impulse Project.

• Other *.exe files are command-line programs. Please refer to the examples in cmd.txt or use the -h option for help.

  • Users can update the corresponding *.int file for each *.exe to perform testing, or execute the program without entering parameters.

Support Functions

NuMaker-M55M1

• Project: The first step in using NuML Studio is to create a project. You can create a project for data collection, ML deployment firmware generation, or both.

• Collecting data: click the - Data Collection tab, and click Download Firmware to EVB to flash firmware into your board.

  • Supports the following options, and users can modify them as needed:
    • G-sensor (3-axis): SDS_Recorder_Gsensor_UART_CMSIS Firmware
    • Audio (16Khz): SDS_Recorder_Audio_UART_CMSIS Firmware
    • Image (supports UVC using the NuMaker-M55M1's image sensor or the PC's webcam): HSUSBH_USBH_UVC Firmware
  • Convert the data to standard format in - Output tab:
    • Users can select csv format to convert sensor *.sds data into *.csv.
    • Users can select audio_wav format to convert audio *.sds data into *.wav.
    • Images collected from Raw Image Collection under Data Collection are already saved as *.jpg

• Upload data to EdgeImpulse:

  • Please update your Edge Impulse project API key in API_key.txt.
  • In - Output tab, user can enable Edge Impulse format & export to upload your data.
  • In - Upload tab, an entire data directory can be uploaded at once.

• Generate ML Model Project:

  • Click the Deployment, - Nuvoton tab, and select your TFLite int8 model. Keil and VSCode CMSIS projects are supported.
    • Model inference firmware code generation.
    • G-sensor model with ARM SDS and RTX5 firmware code generation.
    • Image classification firmware code generation.
    • Object detection firmware code generation.
  • Click the Deployment, - Edge Impulse tab, and select your downloaded Edge Impulse SDK folder (Please select deployment as Ethos-U55-256 library). Keil and VSCode CMSIS projects are supported. (Please update your Edge Impulse project API key in API_key.txt)
    • Model inference firmware code generation.
    • Image classification firmware code generation.
    • KWS firmware code generation.

Tools we use

• NuML_TFLM_Tool: A tool for generating ML model projects based on the TFLM framework. It is based on and references the NuML_TFLM_Tool.
• sds_utilities: A tool for collecting data to a local PC or uploading it to Edge Impulse. Based on the SDS-Framework and Edge Impulse.
• Edge Impulse API: https://docs.edgeimpulse.com/tools/libraries/sdks/studio/python/edgeimpulse/data

Build yourself

• References:

  • Embedded Python Releases for Windows

  • PyStand

• Steps:

  • Download the embedded Python 3.10.6 and copy it to the runtime directory.

  • In a Python virtual environment (venv), use pip to install the dependencies listed in requirements.txt, then copy the site-packages directory to here. Warning: The version of the Python virtual environment must exactly match the embedded Python version.

  • Call the CLI module's entry point directly as a Python module using the embedded Python

    • Edit python310._pth (depends on your version) in runtime to add your site-packages. (The example is in runtime/python310._pth)